This dataset captures the quarterly investment holdings of institutional investment managers and maps the ownership structure of public firms. These Schedule 13F reports are submitted to the Securities and Exchange Commission quarterly by all institutional investment managers with at least $100 million in assets under management. Most academic research examining the common ownership of corporations and the portfolio holdings of large investment managers is based on proprietary commercial databases. This hinders the replication of prior work due to unequal access to these subscriptions and because the data manipulation steps in commercial databases are often opaque. To overcome these limitations, the presented dataset is created from the original regulatory filings; it is updated daily and includes all information reported by investment managers without alteration. Daily updates: https://dx.doi.org/10.34740/kaggle/ds/2973565

Historical SEC dataset containing all insider transactions (Form 4 filings). The data is public, and sourced from the SEC's EDGAR database out of their XML filings. Lightly processed for easier consumption. All Form 4 filings from Jan/20 to Jun/25

Why this exists

Form 4s are noisy to work with: amended filings, multiple insiders per transaction, and inconsistent tables. This dataset provides clean, normalized insider-transaction data with a stable schema so you can backtest signals and monitor insider activity without scraping.

What you get:

• Company (name, ticker if known), CIK
• Insider(s) with role (Director/Officer/10% Owner, etc.)
• Trade details: transaction date, type, number of shares, price, total consideration
• Holdings before/after the transaction

Update cadence: monthly (moving to daily as we scale). Source: U.S. SEC EDGAR Form 4 filings.

We're building a real-time API for new filings with clean JSON endpoints and low latency. If interested, sign up to our waiting list: 👉 https://secfilingapi.com/?utm_source=kaggle&utm_medium=dataset&utm_campaign=form4

Ideas to try with the data:

• Event study: top-decile insider buys vs sector ETF over 30/60/90 days
• Cluster insiders by role + transaction size; test persistence
• Filter for CEO buys after 20% drawdowns (momentum/contrarian mix)
• Find unusual clusters (multiple insiders buying within 10 days)

Limitations & notes:

• There are a few missing filings in the historical data (< 1000 total, all very old filings, before the SEC launched the XML format).
• Forms 4B (amendments to form 4 - rare - are not included)
• Mapping from CIK→ticker can lag for recent IPOs/SPACs.
• Always verify edge cases against the EDGAR link when publishing results.

Feedback & requests welcome in the Discussion tab.

DISCLAIMER: It is possible that inaccuracies or other errors were introduced into the data sets during the process of extracting the data and compiling the data sets. The data set is intended to assist the public in analyzing data contained in Commission filings; however, they are not a substitute for such filings. Investors should review the full Commission filings before making any investment decision.

Browse LSEG's US Company Filings Database, and find a range of filings content and history including annual reports, municipal bonds, and more.

The SEC Form 10-K is an annual report required by the U.S. Securities and Exchange Commission (SEC), that gives a comprehensive summary of a company's financial performance. The full contents of the SEC 10-K are available through the SEC's EDGAR database. PUDL integrates only some of the 10-K metadata and data extracted from the unstructured Exhibit 21 attachement, which describes the ownershp relationships between the parent company and its subsidiaries. This data is used to create a linkage between EIA utilities and SEC reporting companies, to better understand the relationships between utlities and their affiliates, and the resulting economic and political impacts. This data was originally downloaded from the SEC and processed using a machine learning pipeline found here: https://github.com/catalyst-cooperative/mozilla-sec-eia Archived from https://www.sec.gov/search-filings/edgar-application-programming-interfaces

This archive contains raw input data for the Public Utility Data Liberation (PUDL) software developed by Catalyst Cooperative. It is organized into "https://specs.frictionlessdata.io/data-package/">Frictionless Data Packages. For additional information about this data and PUDL, see the following resources:

• The PUDL Repository on GitHub
• PUDL Documentation
• Other Catalyst Cooperative data archives

Material Contracts (Exhibit 10) from SEC/EDGAR

Because sometimes you need 1,141,632 examples of corporate legalese to train your next model ☕

  Dataset Summary

Picture this: 1,141,632 material contracts (Exhibit 10) painstakingly collected from sec.gov's EDGAR database. We're talking about legal agreements spanning from 1994 to 2025 Q1, sourced from 10-K, 10-Q, and 8-K filings. Think of Exhibit 10 as the treasure trove where companies hide their most important legal… See the full description on the dataset page: https://huggingface.co/datasets/chenghao/sec-material-contracts.

LSEG global Filings offers extensive coverage of developed and emerging markets, updated in real time. Discover the data.

Introduction In the course of researching the common ownership hypothesis, we found a number of issues with the Thomson Reuters (TR) "S34" dataset used by many researchers and frequently accessed via Wharton Research Data Services (WRDS). WRDS has done extensive work to improve the database, working with other researchers that have uncovered problems, specifically fixing a lack of records of BlackRock holdings. However, even with the updated dataset posted in the summer of 2018, we discovered a number of discrepancies when accessing data for constituent firms of the S&P 500 Index. We therefore set out to separately create a dataset of 13(f) holdings from the source documents, which are all public and available electronically from the Securities and Exchange Commission (SEC) website. Coverage is good starting in 1999, when electronic filing became mandatory. However, the SEC's Inspector General issued a critical report in 2010 about the information contained in 13(f) filings. The process: We gathered all 13(f) filings from 1999-2017 here. The corpus is over 318,000 filings and occupies ~25GB of space if unzipped. (We do not include the raw filings here as they can be downloaded from EDGAR). We wrote code to parse the filings to extract holding information using regular expressions in Perl. Our target list of holdings was all public firms with a market capitalization of at least $10M. From the header of the file, we first extract the filing date, reporting date, and reporting entity (Central Index Key, or CIK, and CIKNAME). Beginning with the September 30 2013 filing date, all filings were in XML format, which made parsing fairly straightforward, as all values are contained in tags. Prior to that date, the filings are remarkable for the heterogeneity in formatting. Several examples are linked to below. Our approach was to look for any lines containing a CUSIP code that we were interested in, and then attempting to determine the "number of shares" field and the "value" field. To help validate the values we extracted, we downloaded stock price data from CRSP for the filing date, as that allows for a logic check of (price * shares) = value. We do not claim that this will exhaustively extract all holding information. We can provide examples of filings that are formatted in such a way that we are not able to extract the relevant information. In both XML and non-XML filings, we attempt to remove any derivative holdings by looking for phrases such as OPT, CALL, PUT, WARR, etc. We then perform some final data cleaning: in the case of amended filings, we keep an amended level of holdings if the amended report a) occurred within 90 days of the reporting date and b) the initial filing fails our logic check described above. The resulting dataset has around 48M reported holdings (CIK-CUSIP) for all 76 quarters and between 4,000 and 7,000 CUSIPs and between 1,000 and 4,000 investors per quarter. We do not claim that our dataset is perfect; there are undoubtedly errors. As documented elsewhere, there are often errors in the actual source documents as well. However, our method seemed to produce more reliable data in several cases than the TR dataset, as shown in Online Appendix B of the related paper linked above. Included Files Perl Parsing Code (find_holdings_snp.pl). For reference, only needed if you wish to re-parse original filings. Investor holdings for 1999-2017: lightly cleaned. Each CIK-CUSIP-rdate is unique. Over 47M records. The fields are CIK: the central index key assigned by the SEC for this investor. Mapping to names is available below. CUSIP: the identity of the holdings. Consult the SEC's 13(f) listings to identify your CUSIPs of interest. shares: the number of shares reportedly held. Merging in CRSP data on shares outstanding at the CUSIP-Month level allows one to construct \beta. We make no distinction for the sole/shared/none voting discretion fields. If a researcher is interested, we did collect that starting in mid-2013, when filings are in XML format. rdate: reporting date (end of quarter). 8 digit, YYYYMMDD. fdate: filing date. 8 digit, YYYYMMDD. ftype: the form name. Notes: we did not consolidate separate BlackRock entities (or any other possibly related entities). If one wants to do so, use the CIK-CIKname mapping file below. We drop any CUSIP-rdate observation where any investor in that CUSIP reports owning greater than 50% of shares outstanding (even though legitimate cases exist - see, for example, Diamond Offshore and Loews Corporation). We also drop any CUSIP-rdate observation where greater than 120% of shares outstanding are reported to be held by 13(f) investors. Cases where the shares held are listed as zero likely mean the investor filing lists a holding for the firm but that our code could not find the number of shares due to the formatting of the file. We leave these in the data so that any researchers that find a zero know to go back to that source filing to manually gather the holdings for the securities they are interested in. Processed 13f holdings (airlines.parquet, cereal.parquet, out_scrape.parquet). These are used in our related AEJ:Microeconomics paper. The files contain all firms within the airline industry, RTE cereal industry, and all large cap firms (a superset of the S&P 500) respectively. These are a merged version of the scrape_parsed.csv file described above, that include the shares outstanding and percent ownership used to calculate measures of common ownership. These are distributed as brotli compressed Apache Parquet (binary) files. This preserves date information correctly. mgrno: manager number (which is actually CIK in the scraped data) rdate: reporting date ncusip: cusip rrdate: reportaing date in stata format mgrname: manager name shares: shares sole: shares with sole authority shared: shares with shared authority none: shares with no authority isbr/isfi/iss/isba/isvg: is this blackrock, statestreet, vanguard, barclay, fidelity numowners: how many owners prc: price at reporting date shares_out: shares outstanding at reporting date value: reported value in 13(f) beta: shares/shares_out permno: permno Profit weight values (i.e. \kappa) for all firms in the sample. (public_scrape_kappas_XXXX.parquet). Each file represents one year of data and is around 200MB and distributed as a compressed (brotli) parquet file. Fields are simply CUSIP_FROM, CUSIP_TO, KAPPA, QUARTER. Note that these have not been adjusted for multi-class share firms, insider holdings, etc. If looking at a particular market, some additional data cleaning on the investor holdings (above) followed by recomputing profit weights is recommended. For this, we did merge the separate BlackRock entities prior to computing \kappa. CIKmap.csv (~250K observations) Mapping is from CIK-rdate to CIKname. Use this if you want to consolidate holdings across reporting entities or explore the identities of reporting firms. In the case of amended filings that use different names than original ones, we keep the earliest name. Example of Parsing Challenge Prior to the XML era, filings were far from uniform, which creates a notable challenge for parsing them for holdings. In the examples directory we include several example text files of raw 13f filings. Example 1 is a "well behaved" filing, with CUSIP, followed by value, followed by number of shares, as recommended by the SEC. Example 2 shows a case where the ordering is changed: CUSIP, then shares, then value. The column headers show "item 5" coming before "item 4". Example 3 shows a case of a fixed width table, which in principle could be parsed very easily using the tags at the top, although not all filings consistently use these tags. Example 4 shows a case with a fixed width table, with no tag for the CUSIP column. Also, notice that if the firm holds more than 10M shares of a firm, that number occupies the entire width of the column and there is no longer a column separator (i.e. Cisco Systems on line 374). Example 5 shows a comma-separated table format. Example 6 shows a case of changing the column ordering, but also adding an (unrequired) column for share price. Example 7 shows a case where the table is split across subsequent pages, and so the CUSIP appears on a different line than the number of shares.

Dataset contains annotated content from two different data sources: 1) 2,443 sentences from various 2017 SEC contract filings from the publicly available US Securities and Exchange Commission EDGAR (SEC) database, and 2) 21,303 sentences from open source textbooks including topics in biology, history, physics, psychology, economics, sociology, and government.

The Dataset contains structured, parsed data from 758,666 Form 13D and Form 13G blockholder filings from November 1993 to May 2021, downloaded from SEC Edgar. The data is made available as a RAR-compressed CSV file, in which each row represents a single filing and the 76 columns contain parsed information for each filing. Please see the accompanying paper "Determinants of Blockholdership - A new Dataset for Blockholder Analysis" for more information and cite it when using the data for your research. As stated by the SEC, "Information presented on www.sec.gov is considered public information and may be copied or further distributed by users of the web site without the SEC’s permission." (see https://www.sec.gov/privacy.htm#dissemination).

The Corporate Financial Fraud project is a study of company and top-executive characteristics of firms that ultimately violated Securities and Exchange Commission (SEC) financial accounting and securities fraud provisions compared to a sample of public companies that did not. The fraud firm sample was identified through systematic review of SEC accounting enforcement releases from 2005-2010, which included administrative and civil actions, and referrals for criminal prosecution that were identified through mentions in enforcement release, indictments, and news searches. The non-fraud firms were randomly selected from among nearly 10,000 US public companies censused and active during at least one year between 2005-2010 in Standard and Poor's Compustat data. The Company and Top-Executive (CEO) databases combine information from numerous publicly available sources, many in raw form that were hand-coded (e.g., for fraud firms: Accounting and Auditing Enforcement Releases (AAER) enforcement releases, investigation summaries, SEC-filed complaints, litigation proceedings and case outcomes). Financial and structural information on companies for the year leading up to the financial fraud (or around year 2000 for non-fraud firms) was collected from Compustat financial statement data on Form 10-Ks, and supplemented by hand-collected data from original company 10-Ks, proxy statements, or other financial reports accessed via Electronic Data Gathering, Analysis, and Retrieval (EDGAR), SEC's data-gathering search tool. For CEOs, data on personal background characteristics were collected from Execucomp and BoardEx databases, supplemented by hand-collection from proxy-statement biographies.

The MarkupMnA dataset is a corpus of 151 merger and acquisition agreements with annotated sections titles, section numbers, page numbers, and more, based on HTML filings by US public companies retrieved from the SEC EDGAR database. We consider the task of section title annotation as a sequence labeling task, and to that end, use the BEIOS tagging scheme when generating our annotations. There are over 70,000 labels in the entire dataset excluding outside labels and over 465,000 labels including outside labels.

We add annotations to the contracts in an already widely used dataset, MAUD, which is an expert-annotated reading comprehension dataset. The broad objective of our work is to make progress toward developing computationally efficient hierarchical representations of long documents, specifically for legal contracts. We hope that our annotations can be used in conjunction with MAUD to advance legal NLP research.

Files include: (1) an open sourced database of CEO duality and board chair orientations developed by scaling human coded data using supervised machine learning techniques (in both .dta and .csv formats), as well as (2) the accompanying training and scoring scripts to scale human coded data.

Users may apply the scoring script to score the same variables from company proxy statements, or may adapt the training/scoring scripts and retrain models to scale human coded data of other constructs or measures.

We note that early steps in the process to develop our database and script required web-scraping of company filings from SEC Edgar and text extraction from collected filings. We relied on other publicly available scripts to develop our own fetcher and extraction scripts. Users seeking to duplicate those parts of the process may benefit from the following resources from Kai Chen and pipy.org:

For resources from Kai Chen: see https://www.kaichen.work/?p=681 and https://www.kaichen.work/?p=946

For resources from pipy.org, see sec-edgar-downloader and sec-api

The DEFT corpus1 consists of annotated content from two different data sources: 1) 2,443 sentences (5,324,430 tokens) from various 2017 SEC contract filings from the publicly available US Securities and Exchange Commission EDGAR (SEC) database, and 2) 21,303 sentences (409,253 tokens) from the https://cnx.org/ open source textbooks (by various authors, licensed under CC BY 4.0) including topics in biology, history, physics, psychology, economics, sociology, and government. 22% of SEC sentences contain definitions and 28% of textbook sentences contain definitions. Our entire corpus, including both datasets, is significantly larger and more complex than any existing definition extraction dataset (see Table 1).